from src.algorithms.classic_region_growing import ClassicRegionGrowing

ClassicRegionGrowing??

Init signature:
ClassicRegionGrowing(
    threshold: float = 0.1,
    min_region_size: int = 50,
    cloud_mask_value: float = -999.0,
)
Source:        
class ClassicRegionGrowing:
    """
    Classic Region Growing algorithm for spectral-based image segmentation.

    This implementation uses BFS with 4-connectivity to grow regions from seed points
    based on a homogeneity criterion (spectral similarity threshold).

    Attributes:
        threshold: Maximum spectral difference to consider pixels as similar
        min_region_size: Minimum number of pixels for a valid region
        cloud_mask_value: Sentinel value for masked pixels (clouds, no-data)
    """

    def __init__(
        self,
        threshold: float = 0.1,
        min_region_size: int = 50,
        cloud_mask_value: float = -999.0
    ):
        """
        Initialize the Region Growing algorithm.

        Args:
            threshold: Homogeneity threshold for spectral similarity.
                      For NDVI: |NDVI_A - NDVI_B| < threshold
                      Typical values: 0.05-0.15 (default: 0.1)
                      Lower values = more strict, more regions
                      Higher values = more permissive, fewer regions
            min_region_size: Minimum region size in pixels to filter noise.
                            Typical values: 25-100 (default: 50)
                            At 10m resolution: 50 pixels = 5000 m² = 0.5 hectares
            cloud_mask_value: Value indicating masked pixels (clouds, no-data).
                             Default: -999.0 (sentinel value)
        """
        if threshold <= 0:
            raise ValueError("Threshold must be positive")
        if min_region_size < 1:
            raise ValueError("Minimum region size must be at least 1")

        self.threshold = threshold
        self.min_region_size = min_region_size
        self.cloud_mask_value = cloud_mask_value

    def segment(
        self,
        image: np.ndarray,
        seeds: Optional[List[Tuple[int, int]]] = None
    ) -> Tuple[np.ndarray, int, List[Dict]]:
        """
        Segment image using Region Growing algorithm.

        Args:
            image: 2D NumPy array with spectral values (e.g., NDVI).
                  Expected range: [-1, 1] for NDVI
                  Masked pixels should have value == cloud_mask_value
            seeds: List of seed coordinates [(y, x), ...].
                  If None, seeds are generated automatically using grid method.

        Returns:
            Tuple containing:
            - labeled_image: 2D array with region labels (0=no region, 1..N=region IDs)
            - num_regions: Total number of valid regions found
            - regions_info: List of dicts with information for each region:
                * id: Region ID (1-indexed)
                * size: Number of pixels
                * mean_ndvi: Mean spectral value
                * std_ndvi: Standard deviation
                * min_ndvi: Minimum value
                * max_ndvi: Maximum value
                * pixels: List of (y, x) coordinates

        Example:
            >>> algorithm = ClassicRegionGrowing(threshold=0.1, min_region_size=50)
            >>> ndvi_image = np.random.rand(256, 256)
            >>> labeled, num_regions, info = algorithm.segment(ndvi_image)
            >>> print(f"Found {num_regions} regions")
        """
        if image.ndim != 2:
            raise ValueError(f"Image must be 2D, got shape {image.shape}")

        h, w = image.shape
        labeled = np.zeros((h, w), dtype=np.int32)
        region_id = 0
        regions_info = []

        # Generate seeds if not provided
        if seeds is None:
            seeds = self.generate_seeds_grid(image)

        # Grow region from each seed
        for seed_y, seed_x in seeds:
            # Skip if already labeled
            if labeled[seed_y, seed_x] != 0:
                continue

            # Skip if masked (cloud/no-data)
            if self._is_masked(image[seed_y, seed_x]):
                continue

            # Grow region from this seed
            region_pixels = self._grow_region(image, seed_y, seed_x, labeled, region_id + 1)

            # Validate region size
            if len(region_pixels) >= self.min_region_size:
                region_id += 1

                # Calculate region statistics
                region_values = [image[y, x] for y, x in region_pixels]
                region_info = {
                    'id': region_id,
                    'size': len(region_pixels),
                    'mean_ndvi': float(np.mean(region_values)),
                    'std_ndvi': float(np.std(region_values)),
                    'min_ndvi': float(np.min(region_values)),
                    'max_ndvi': float(np.max(region_values)),
                    'pixels': region_pixels
                }
                regions_info.append(region_info)
            else:
                # Region too small, remove labels
                for y, x in region_pixels:
                    labeled[y, x] = 0

        return labeled, region_id, regions_info

    def _grow_region(
        self,
        image: np.ndarray,
        start_y: int,
        start_x: int,
        labeled: np.ndarray,
        region_id: int
    ) -> List[Tuple[int, int]]:
        """
        Grow a region from a seed point using BFS (Breadth-First Search).

        This method implements the core Region Growing algorithm using BFS with 4-connectivity.
        It expands from the seed point by adding neighboring pixels that satisfy the
        homogeneity criterion.

        Args:
            image: 2D array with spectral values
            start_y: Y-coordinate of seed point
            start_x: X-coordinate of seed point
            labeled: 2D array to store region labels (modified in-place)
            region_id: ID to assign to this region

        Returns:
            List of pixel coordinates [(y, x), ...] belonging to the region

        Algorithm:
            1. Initialize queue with seed point
            2. While queue not empty:
                a. Pop pixel from queue
                b. If valid and similar to seed: add to region
                c. Add 4-connected neighbors to queue
            3. Return region pixels
        """
        h, w = image.shape
        seed_value = image[start_y, start_x]

        # BFS queue and tracking
        queue = deque([(start_y, start_x)])
        region_pixels = []
        visited = set()

        while queue:
            y, x = queue.popleft()

            # Skip if already visited
            if (y, x) in visited:
                continue

            # Check bounds
            if not (0 <= y < h and 0 <= x < w):
                continue

            # Skip if already labeled
            if labeled[y, x] != 0:
                continue

            # Skip masked pixels
            pixel_value = image[y, x]
            if self._is_masked(pixel_value):
                continue

            # Check homogeneity criterion
            if abs(pixel_value - seed_value) > self.threshold:
                continue

            # Add pixel to region
            visited.add((y, x))
            region_pixels.append((y, x))
            labeled[y, x] = region_id

            # Add 4-connected neighbors to queue
            queue.append((y + 1, x))  # Down
            queue.append((y - 1, x))  # Up
            queue.append((y, x + 1))  # Right
            queue.append((y, x - 1))  # Left

        return region_pixels

    def generate_seeds_grid(
        self,
        image: np.ndarray,
        grid_size: int = 20
    ) -> List[Tuple[int, int]]:
        """
        Generate seed points using a regular grid pattern.

        This method creates a grid of evenly spaced seed points across the image.
        Only valid (non-masked) pixels are selected as seeds.

        Args:
            image: 2D array with spectral values
            grid_size: Spacing between seeds in pixels (default: 20)
                      Smaller values = more seeds = finer segmentation
                      Larger values = fewer seeds = coarser segmentation

        Returns:
            List of seed coordinates [(y, x), ...]

        Example:
            For a 512x512 image with grid_size=20:
            - Seeds will be placed at (10, 10), (10, 30), ..., (10, 510)
            - Total seeds ≈ (512/20)² ≈ 650 seeds
        """
        h, w = image.shape
        seeds = []

        # Create grid starting from half grid_size offset
        for y in range(grid_size // 2, h, grid_size):
            for x in range(grid_size // 2, w, grid_size):
                # Only add valid pixels as seeds
                pixel_value = image[y, x]
                if not self._is_masked(pixel_value):
                    seeds.append((y, x))

        return seeds

    def classify_by_stress(
        self,
        regions_info: List[Dict]
    ) -> Dict[str, List[Dict]]:
        """
        Classify regions by vegetation stress level based on mean NDVI.

        NDVI stress level thresholds (standard values for crops):
        - High stress: NDVI < 0.3 (bare soil, senescent vegetation, severe stress)
        - Medium stress: 0.3 ≤ NDVI < 0.5 (moderate vegetation, moderate stress)
        - Low stress: NDVI ≥ 0.5 (healthy vegetation, low/no stress)

        Args:
            regions_info: List of region information dicts (from segment method)

        Returns:
            Dict with classified regions:
            {
                'high_stress': [region1, region2, ...],
                'medium_stress': [region3, region4, ...],
                'low_stress': [region5, region6, ...]
            }
            Each region dict gets an additional 'stress_level' field.
        """
        classified = {
            'high_stress': [],
            'medium_stress': [],
            'low_stress': []
        }

        for region in regions_info:
            mean_ndvi = region['mean_ndvi']

            if mean_ndvi < 0.3:
                region['stress_level'] = 'high'
                classified['high_stress'].append(region)
            elif mean_ndvi < 0.5:
                region['stress_level'] = 'medium'
                classified['medium_stress'].append(region)
            else:
                region['stress_level'] = 'low'
                classified['low_stress'].append(region)

        return classified

    def _is_masked(self, value: float) -> bool:
        """
        Check if a pixel value represents a masked area (cloud, no-data).

        Args:
            value: Pixel spectral value

        Returns:
            True if pixel is masked, False otherwise
        """
        # Check for NaN, Inf, and cloud mask value
        if np.isnan(value) or np.isinf(value):
            return True
        if abs(value - self.cloud_mask_value) < 1e-6:
            return True
        return False
File:           c:\users\arthu\proyectos\mna\region-growing\src\algorithms\classic_region_growing.py
Type:           type
Subclasses:

import torch
import sys

print("VERIFICACIÓN DE ENTORNO")
print(f"Python: {sys.version}")
print(f"PyTorch: {torch.__version__}")
print(f"CUDA available: {torch.cuda.is_available()}")
if torch.cuda.is_available():
    print(f"GPU: {torch.cuda.get_device_name(0)}")
    print(f"GPU memory: {torch.cuda.get_device_properties(0).total_memory / 1e9:.2f} GB")
else:
    print("⚠️ WARNING: No GPU detected")

VERIFICACIÓN DE ENTORNO
Python: 3.12.6 (tags/v3.12.6:a4a2d2b, Sep  6 2024, 20:11:23) [MSC v.1940 64 bit (AMD64)]
PyTorch: 2.9.0+cu129
CUDA available: True
GPU: NVIDIA GeForce RTX 4070 Laptop GPU
GPU memory: 8.59 GB

# Standard library
import sys
sys.path.append('../..')
import time
from pathlib import Path
import os
import json
import time
import warnings
from pathlib import Path

# Data processing
import numpy as np
import pandas as pd

# Visualization
import matplotlib.pyplot as plt
import matplotlib.patches as mpatches
import seaborn as sns

# Machine Learning
import torch


# Project imports

from src.features.hls_processor import (
    extract_embeddings,
    save_embeddings,
    load_embeddings,
)

from src.utils.visualization_utils import (
    visualize_multispectral_zones,
    visualize_segmentation_results,
    visualize_method_comparison,
    visualize_ab_comparison_metrics,
    visualize_dynamic_world_comparison,
    visualize_semantic_maps,
    visualize_class_statistics,
    visualize_class_distribution_summary,
    visualize_class_statistics_all_zones,
    analyze_ndvi_zones
)


from src.classification.zero_shot_classifier import (
    LAND_COVER_CLASSES,
    CLASS_COLORS
)

from src.utils.test_zones import MEXICO_TEST_ZONES, load_or_download_zones
from src.features.ndvi_calculator import calculate_ndvi

from src.utils.segmentation_utils import (
    run_classic_rg_segmentation,
    classify_stress_levels,
    calculate_quality_metrics,
    create_comprehensive_summary,
    run_mgrg_segmentation,
    compare_segmentation_results,
    print_comparison_summary,
    create_detailed_method_comparison,
)

from src.utils.validation_utils import (
    VALIDATION_CLASSES,
    print_validation_classes,
    validate_all_zones,
    calculate_aggregate_metrics,
    analyze_relative_improvements,
    cross_validate_with_dynamic_world,
    
    
    
)

from src.utils.classification_utils import (
    analyze_crop_stress,
    classify_all_zones,

)

# Flat
RASTERIO_AVAILABLE = True

%matplotlib inline

# Configure
warnings.filterwarnings('ignore')
plt.style.use('seaborn-v0_8-darkgrid')
plt.rcParams['figure.dpi'] = 300
plt.rcParams['font.size'] = 12
plt.rcParams['figure.figsize'] = (16, 10)

# Add src to path
project_root = Path.cwd().parent.parent
sys.path.append(str(project_root))

print("All Imports successful")
print(f"Project root: {project_root}")

All Imports successful
Project root: c:\Users\arthu\Proyectos\MNA\region-growing

# Define Mexican test zones
pd.set_option('display.max_colwidth', None)
zonas_df = pd.DataFrame.from_dict(MEXICO_TEST_ZONES, orient='index')
zonas_df.index.name = 'Zona ID'

# Display zones information
display(zonas_df)

# Configuration parameters
DATE_FROM = '2025-10-10'
DATE_TO = '2025-11-10'
USE_SIMPLE_MODEL = False  # Set to True for testing without full Prithvi download

# Output directories
OUTPUT_DIR = project_root / 'img' / 'sentinel2' / 'embeddings'
OUTPUT_DIR.mkdir(parents=True, exist_ok=True)

print(f"Fecha de análisis: {DATE_FROM}")
print(f"Modelo: {'Simple (desarrollo)' if USE_SIMPLE_MODEL else 'Prithvi (producción)'}")
print(f"Directorio de salida: {OUTPUT_DIR}")

Fecha de análisis: 2025-10-10
Modelo: Prithvi (producción)
Directorio de salida: c:\Users\arthu\Proyectos\MNA\region-growing\img\sentinel2\embeddings

zones_data = load_or_download_zones(
    date_from=DATE_FROM,
    date_to=DATE_TO,
    project_root=project_root,
    sentinel_client_id=os.getenv('SENTINEL_CLIENT_ID'),
    sentinel_client_secret=os.getenv('SENTINEL_CLIENT_SECRET')
)

Processing Valle de Mexicali...
  Loading existing HLS image from c:\Users\arthu\Proyectos\MNA\region-growing\img\sentinel2\mexico\mexicali\hls_image.npy
  Loaded Valle de Mexicali: (6, 1124, 922)

Processing El Bajío...
  Loading existing HLS image from c:\Users\arthu\Proyectos\MNA\region-growing\img\sentinel2\mexico\bajio\hls_image.npy
  Loaded El Bajío: (6, 1092, 1057)

Processing Valle de Culiacán...
  Loading existing HLS image from c:\Users\arthu\Proyectos\MNA\region-growing\img\sentinel2\mexico\sinaloa\hls_image.npy
  Loaded Valle de Culiacán: (6, 1090, 1031)

Processing complete: 3/3 zones loaded

# Configuración
OUTPUT_DIR = Path('../../img/results/classification')
OUTPUT_DIR.mkdir(parents=True, exist_ok=True)

# Visualizar todas las zonas en una sola figura
fig = visualize_multispectral_zones(
    zones_data=zones_data,
    output_path=OUTPUT_DIR / 'multispectral_composites.png',
    figsize=(18, 16),
    dpi=150,
    show_plot=True
)

Visualización guardada en: ..\..\img\results\classification\multispectral_composites.png

# SECCIÓN: IMPLEMENTACIÓN DEL ALGORITMO CLASSIC REGION GROWING

# Configuración del algoritmo
THRESHOLD = 0.1  # Umbral de similitud NDVI
MIN_REGION_SIZE = 50  # Tamaño mínimo de región en píxeles
SEED_SPACING = 20  # Espaciado de semillas en píxeles

print("ALGORITMO: CLASSIC REGION GROWING")
print("Parámetros:")
print(f"  - Threshold: {THRESHOLD}")
print(f"  - Min region size: {MIN_REGION_SIZE} píxeles")
print(f"  - Seed spacing: {SEED_SPACING} píxeles")

# Procesar cada zona
results_classic = {}

ALGORITMO: CLASSIC REGION GROWING
Parámetros:
  - Threshold: 0.1
  - Min region size: 50 píxeles
  - Seed spacing: 20 píxeles

ndvi_results = {}

for zone_id, zone_data in zones_data.items():
    if zone_data is None:
        continue
    
    zone_name = zone_data['config']['name']
    print(f"\nProcesando: {zone_name}")
    
    hls_image = zone_data['hls_image']
    
    # Calcular NDVI
    red_band = hls_image[2]  # B04
    nir_band = hls_image[3]  # B8A
    
    ndvi_result = calculate_ndvi(red_band, nir_band)
    ndvi = ndvi_result['ndvi_masked']  # ← CORREGIDO
    
    ndvi_results[zone_id] = ndvi
    
    print("  ✓ NDVI calculado")
    print(f"    - Rango: [{ndvi.min():.3f}, {ndvi.max():.3f}]")
    print(f"    - Media: {ndvi.mean():.3f}")
    print(f"    - Desviación estándar: {ndvi.std():.3f}")

print("NDVI calculado para todas las zonas")

Procesando: Valle de Mexicali
  ✓ NDVI calculado
    - Rango: [-0.580, 0.913]
    - Media: 0.191
    - Desviación estándar: 0.141

Procesando: El Bajío
  ✓ NDVI calculado
    - Rango: [-0.381, 0.888]
    - Media: 0.315
    - Desviación estándar: 0.132

Procesando: Valle de Culiacán
  ✓ NDVI calculado
    - Rango: [-0.493, 0.853]
    - Media: 0.198
    - Desviación estándar: 0.195
NDVI calculado para todas las zonas

# GENERACIÓN DE SEMILLAS

print("GENERACIÓN DE SEMILLAS")
print(f"Configuración: Grid {SEED_SPACING}x{SEED_SPACING} píxeles\n")

seeds_by_zone = {}

for zone_id, zone_data in zones_data.items():
    if zone_data is None:
        continue
    
    zone_name = zone_data['config']['name']
    hls_image = zone_data['hls_image']
    h, w = hls_image.shape[1], hls_image.shape[2]
    
    # Generar semillas en grid
    seeds = []
    for y in range(SEED_SPACING, h, SEED_SPACING):
        for x in range(SEED_SPACING, w, SEED_SPACING):
            seeds.append((y, x))
    
    seeds_by_zone[zone_id] = seeds
    
    print(f"{zone_name}:")
    print(f"  - Dimensiones: {h}x{w} píxeles")
    print(f"  - Semillas generadas: {len(seeds)}")

print("Semillas generadas para todas las zonas")

GENERACIÓN DE SEMILLAS
Configuración: Grid 20x20 píxeles

Valle de Mexicali:
  - Dimensiones: 1124x922 píxeles
  - Semillas generadas: 2576
El Bajío:
  - Dimensiones: 1092x1057 píxeles
  - Semillas generadas: 2808
Valle de Culiacán:
  - Dimensiones: 1090x1031 píxeles
  - Semillas generadas: 2754
Semillas generadas para todas las zonas

# Ejecutar segmentación Classic RG
segmentation_results = run_classic_rg_segmentation(
    zones_data=zones_data,
    ndvi_results=ndvi_results,
    seeds_by_zone=seeds_by_zone,
    threshold=THRESHOLD,
    min_region_size=MIN_REGION_SIZE
)

================================================================================
SEGMENTACIÓN CON CLASSIC REGION GROWING
================================================================================

Parámetros:
  • Threshold (NDVI): 0.1
  • Min region size: 50 píxeles
--------------------------------------------------------------------------------

Segmentando: Valle de Mexicali
  ✓ Completado en 2.70s
    - Regiones: 300
    - Cobertura: 85.0%
    - Velocidad: 0.38 M píxeles/s

Segmentando: El Bajío
  ✓ Completado en 2.99s
    - Regiones: 451
    - Cobertura: 79.9%
    - Velocidad: 0.39 M píxeles/s

Segmentando: Valle de Culiacán
  ✓ Completado en 2.27s
    - Regiones: 632
    - Cobertura: 55.8%
    - Velocidad: 0.50 M píxeles/s

================================================================================
✓ SEGMENTACIÓN COMPLETADA PARA 3 ZONAS
================================================================================

classic_stress_classification = classify_stress_levels(
    segmentation_results=segmentation_results,
    ndvi_results=ndvi_results,
    zones_data=zones_data
)

================================================================================
CLASIFICACIÓN POR NIVEL DE ESTRÉS
================================================================================

Criterios (basados en NDVI):
  • Alto:   NDVI < 0.3
  • Medio:  0.3 ≤ NDVI < 0.5
  • Bajo:   NDVI ≥ 0.5
--------------------------------------------------------------------------------

Clasificando: Valle de Mexicali
  ✓ Clasificación completada
    - Estrés alto:   142 regiones
    - Estrés medio:  107 regiones
    - Estrés bajo:   51 regiones

Clasificando: El Bajío
  ✓ Clasificación completada
    - Estrés alto:   241 regiones
    - Estrés medio:  123 regiones
    - Estrés bajo:   87 regiones

Clasificando: Valle de Culiacán
  ✓ Clasificación completada
    - Estrés alto:   356 regiones
    - Estrés medio:  167 regiones
    - Estrés bajo:   109 regiones

================================================================================
RESUMEN DE CLASIFICACIÓN POR ESTRÉS
================================================================================

================================================================================
✓ Clasificación por estrés completada para 3 zonas

# Calcular métricas para Classic RG
classic_quality_metrics, df_classic = calculate_quality_metrics(
    segmentation_results=segmentation_results,
    ndvi_results=ndvi_results,
    zones_data=zones_data
)

====================================================================================================
MÉTRICAS DE CALIDAD DE SEGMENTACIÓN
====================================================================================================

====================================================================================================

# 2. Resumen completo de Classic RG
df_classic_summary, classic_insights = create_comprehensive_summary(
    segmentation_results=segmentation_results,
    quality_metrics=classic_quality_metrics,
    stress_classification=classic_stress_classification,
    zones_data=zones_data,
    method_name="Classic RG"
)

========================================================================================================================
RESUMEN COMPARATIVO ENTRE ZONAS - Classic RG
========================================================================================================================

------------------------------------------------------------------------------------------------------------------------
INSIGHTS CLAVE:
------------------------------------------------------------------------------------------------------------------------
  • Coherencia espacial promedio:      95.8%
  • Cobertura promedio:                 73.6%
  • Regiones promedio por zona:         461
  • Tamaño promedio de región:          19.91 hectáreas
  • Tiempo promedio de procesamiento:   2.58 segundos

  • Total regiones con estrés alto:     739
  • Total regiones con estrés medio:    397
  • Total regiones con estrés bajo:     247
========================================================================================================================

fig_classic = visualize_segmentation_results(
    zones_data=zones_data,
    ndvi_results=ndvi_results,
    segmentation_results=segmentation_results,
    stress_classification=classic_stress_classification,
    output_path=OUTPUT_DIR / 'classic_rg_complete_results.png'
)

Visualización completa guardada en: ..\..\img\results\classification\classic_rg_complete_results.png

# Extract embeddings for all zones
for zone_id, zone_data in zones_data.items():
    if zone_data is None:
        continue
        
    print(f"\nExtrayendo embeddings de {zone_data['config']['name']}...")
    
    try:
        # Extract embeddings
        embeddings = extract_embeddings(
            hls_image=zone_data['hls_image'],
            use_simple_model=USE_SIMPLE_MODEL
        )
        
        # Save embeddings
        output_path = OUTPUT_DIR / f"{zone_id}_{DATE_FROM.replace('-', '')}.npz"
        metadata = {
            'zone': zone_id,
            'zone_name': zone_data['config']['name'],
            'date': DATE_FROM,
            'bbox': zone_data['config']['bbox'],
            'model': 'simple' if USE_SIMPLE_MODEL else 'prithvi'
        }
        
        save_embeddings(embeddings, output_path, metadata)
        
        # Store in dictionary
        zones_data[zone_id]['embeddings'] = embeddings
        zones_data[zone_id]['embeddings_path'] = output_path
        
        # Verify L2 normalization
        sample_norm = np.linalg.norm(embeddings[0, 0])
        
        print(f"  Shape: {embeddings.shape}")
        print(f"  L2 norm (sample pixel): {sample_norm:.6f}")
        print(f"  Guardado en: {output_path.name}")
        
    except Exception as e:
        print(f"  Error: {e}")
        zones_data[zone_id]['embeddings'] = None

print("\nExtracción de embeddings completa")

Extrayendo embeddings de Mexicali...
  Shape: (1124, 922, 256)
  L2 norm (sample pixel): 1.000000
  Guardado en: mexicali_20251010.npz

Extrayendo embeddings de Bajío...
  Shape: (1092, 1057, 256)
  L2 norm (sample pixel): 1.000000
  Guardado en: bajio_20251010.npz

Extrayendo embeddings de Sinaloa...
  Shape: (1090, 1031, 256)
  L2 norm (sample pixel): 1.000000
  Guardado en: sinaloa_20251010.npz

Extracción de embeddings completa

embeddings_dir = Path('../../img/sentinel2/embeddings')

zones = {
    'mexicali': 'mexicali_20251010.npz',
    'bajio': 'bajio_20251010.npz',
    'sinaloa': 'sinaloa_20251010.npz'
}

embeddings_data = {}

for zone_name, filename in zones.items():
    filepath = embeddings_dir / filename
    if filepath.exists():
        embeddings, metadata = load_embeddings(filepath)
        embeddings_data[zone_name] = {
            'embeddings': embeddings,
            'metadata': metadata
        }
        print(f"{zone_name.capitalize()}: {embeddings.shape}")
    else:
        print(f"Warning: {filepath} not found")

if len(embeddings_data) == 0:
    print("\nNo embeddings found. Creating synthetic embeddings for demo...")
    np.random.seed(42)
    embeddings_demo = np.random.rand(256, 256, 256)
    embeddings_demo = embeddings_demo / np.linalg.norm(embeddings_demo, axis=2, keepdims=True)
    embeddings_data['demo'] = {
        'embeddings': embeddings_demo,
        'metadata': {'zone': 'demo', 'synthetic': True}
    }
    print(f"Demo: {embeddings_demo.shape}")

Mexicali: (1124, 922, 256)
Bajio: (1092, 1057, 256)
Sinaloa: (1090, 1031, 256)

# IMPLEMENTACIÓN DE MGRG (REGION GROWING SEMÁNTICO)
from src.utils.segmentation_utils import run_mgrg_segmentation

# Parámetros de configuración
THRESHOLD_MGRG = 0.85  # Cosine similarity threshold
MIN_REGION_SIZE = 50

# Mapeo de nombres de embeddings a zone_ids
zone_mapping = {
    'mexicali': 'mexicali',
    'bajio': 'bajio',
    'sinaloa': 'sinaloa'
}

# Ejecutar segmentación MGRG
mgrg_results = run_mgrg_segmentation(
    zones_data=zones_data,
    embeddings_data=embeddings_data,
    threshold=THRESHOLD_MGRG,
    min_region_size=MIN_REGION_SIZE,
    seeds_by_zone=seeds_by_zone,  # Opcional: reutilizar semillas del Classic RG
    segmentation_results=segmentation_results,  # Opcional: para comparación
    zone_mapping=zone_mapping
)

================================================================================
SEGMENTACIÓN CON MGRG (METRIC-GUIDED REGION GROWING)
================================================================================

Parámetros:
  • Threshold (cosine similarity): 0.85
  • Min region size: 50 píxeles
  • Criterio: Similitud semántica en espacio de embeddings
--------------------------------------------------------------------------------

Segmentando: Valle de Mexicali
  Embeddings shape: (1124, 922, 256)
  Semillas (reutilizadas): 2576
  Ejecutando MGRG...
  ✓ Completado en 3.08s
    - Regiones MGRG: 137
    - Regiones Classic RG: 300
    - Reducción: 54.3%
    - Cobertura: 98.6%

Segmentando: El Bajío
  Embeddings shape: (1092, 1057, 256)
  Semillas (reutilizadas): 2808
  Ejecutando MGRG...
  ✓ Completado en 3.35s
    - Regiones MGRG: 297
    - Regiones Classic RG: 451
    - Reducción: 34.1%
    - Cobertura: 95.8%

Segmentando: Valle de Culiacán
  Embeddings shape: (1090, 1031, 256)
  Semillas (reutilizadas): 2754
  Ejecutando MGRG...
  ✓ Completado en 2.65s
    - Regiones MGRG: 1181
    - Regiones Classic RG: 632
    - Reducción: -86.9%
    - Cobertura: 89.0%

================================================================================
✓ MGRG COMPLETADO PARA 3 ZONAS
================================================================================

# Configuración de parámetros
THRESHOLD_MGRG = 0.85
MIN_REGION_SIZE = 50

# Mapeo de nombres de embeddings a zone_ids
zone_mapping = {
    'mexicali': 'mexicali',
    'bajio': 'bajio',
    'sinaloa': 'sinaloa'
}

# Ejecutar MGRG
mgrg_results = run_mgrg_segmentation(
    zones_data=zones_data,
    embeddings_data=embeddings_data,
    threshold=0.85,
    min_region_size=50,
    seeds_by_zone=seeds_by_zone,
    segmentation_results=segmentation_results,
    zone_mapping=zone_mapping
)

# Comparar resultados
comparison = compare_segmentation_results(
    classic_results=segmentation_results,
    mgrg_results=mgrg_results,
    zones_data=zones_data
)

print_comparison_summary(comparison)

================================================================================
SEGMENTACIÓN CON MGRG (METRIC-GUIDED REGION GROWING)
================================================================================

Parámetros:
  • Threshold (cosine similarity): 0.85
  • Min region size: 50 píxeles
  • Criterio: Similitud semántica en espacio de embeddings
--------------------------------------------------------------------------------

Segmentando: Valle de Mexicali
  Embeddings shape: (1124, 922, 256)
  Semillas (reutilizadas): 2576
  Ejecutando MGRG...
  ✓ Completado en 2.87s
    - Regiones MGRG: 137
    - Regiones Classic RG: 300
    - Reducción: 54.3%
    - Cobertura: 98.6%

Segmentando: El Bajío
  Embeddings shape: (1092, 1057, 256)
  Semillas (reutilizadas): 2808
  Ejecutando MGRG...
  ✓ Completado en 3.38s
    - Regiones MGRG: 297
    - Regiones Classic RG: 451
    - Reducción: 34.1%
    - Cobertura: 95.8%

Segmentando: Valle de Culiacán
  Embeddings shape: (1090, 1031, 256)
  Semillas (reutilizadas): 2754
  Ejecutando MGRG...
  ✓ Completado en 2.47s
    - Regiones MGRG: 1181
    - Regiones Classic RG: 632
    - Reducción: -86.9%
    - Cobertura: 89.0%

================================================================================
✓ MGRG COMPLETADO PARA 3 ZONAS
================================================================================


====================================================================================================
COMPARACIÓN: CLASSIC RG vs MGRG
====================================================================================================

----------------------------------------------------------------------------------------------------
ESTADÍSTICAS GENERALES:
----------------------------------------------------------------------------------------------------
  Promedio de regiones Classic RG: 461
  Promedio de regiones MGRG:       538
  Reducción promedio:              0.5%
====================================================================================================

# Clasificar estrés para MGRG
mgrg_stress_classification = classify_stress_levels(
    segmentation_results=mgrg_results,
    ndvi_results=ndvi_results,
    zones_data=zones_data
)

================================================================================
CLASIFICACIÓN POR NIVEL DE ESTRÉS
================================================================================

Criterios (basados en NDVI):
  • Alto:   NDVI < 0.3
  • Medio:  0.3 ≤ NDVI < 0.5
  • Bajo:   NDVI ≥ 0.5
--------------------------------------------------------------------------------

Clasificando: Valle de Mexicali
  ✓ Clasificación completada
    - Estrés alto:   95 regiones
    - Estrés medio:  27 regiones
    - Estrés bajo:   15 regiones

Clasificando: El Bajío
  ✓ Clasificación completada
    - Estrés alto:   263 regiones
    - Estrés medio:  29 regiones
    - Estrés bajo:   5 regiones

Clasificando: Valle de Culiacán
  ✓ Clasificación completada
    - Estrés alto:   1051 regiones
    - Estrés medio:  121 regiones
    - Estrés bajo:   9 regiones

================================================================================
RESUMEN DE CLASIFICACIÓN POR ESTRÉS
================================================================================

================================================================================
✓ Clasificación por estrés completada para 3 zonas

# Calcular métricas para MGRG (solo muestra DataFrame, sin prints)
mgrg_quality_metrics, df_mgrg = calculate_quality_metrics(
    segmentation_results=mgrg_results,
    ndvi_results=ndvi_results,
    zones_data=zones_data
)

====================================================================================================
MÉTRICAS DE CALIDAD DE SEGMENTACIÓN
====================================================================================================

====================================================================================================

# Resumen completo de MGRG
df_mgrg_summary, mgrg_insights = create_comprehensive_summary(
    segmentation_results=mgrg_results,
    quality_metrics=mgrg_quality_metrics,
    stress_classification=mgrg_stress_classification,
    zones_data=zones_data,
    method_name="MGRG"
)

========================================================================================================================
RESUMEN COMPARATIVO ENTRE ZONAS - MGRG
========================================================================================================================

------------------------------------------------------------------------------------------------------------------------
INSIGHTS CLAVE:
------------------------------------------------------------------------------------------------------------------------
  • Coherencia espacial promedio:      88.4%
  • Cobertura promedio:                 94.5%
  • Regiones promedio por zona:         538
  • Tamaño promedio de región:          40.09 hectáreas
  • Tiempo promedio de procesamiento:   2.91 segundos

  • Total regiones con estrés alto:     1409
  • Total regiones con estrés medio:    177
  • Total regiones con estrés bajo:     29
========================================================================================================================

# 6. Visualizar MGRG
fig_mgrg = visualize_segmentation_results(
    zones_data=zones_data,
    ndvi_results=ndvi_results,
    segmentation_results=mgrg_results,
    stress_classification=mgrg_stress_classification,
    output_path=OUTPUT_DIR / 'mgrg_complete_results.png'
)

Visualización completa guardada en: ..\..\img\results\classification\mgrg_complete_results.png

# Comparación detallada con filas de mejora intercaladas
df_detailed_comparison = create_detailed_method_comparison(
    classic_results=segmentation_results,
    classic_quality=classic_quality_metrics,
    classic_stress=classic_stress_classification,
    mgrg_results=mgrg_results,
    mgrg_quality=mgrg_quality_metrics,
    mgrg_stress=mgrg_stress_classification,
    zones_data=zones_data
)

# Acceder a insights programáticamente
print(f"\nINSIGHTS:")
print(f"  MGRG tiene {mgrg_insights['avg_coherence']:.1f}% de coherencia")
print(f"  Classic RG tiene {classic_insights['avg_coherence']:.1f}% de coherencia")
print(f"  Mejora: {mgrg_insights['avg_coherence'] - classic_insights['avg_coherence']:+.1f}%")

========================================================================================================================
RESUMEN COMPARATIVO DETALLADO: CLASSIC RG vs MGRG
========================================================================================================================

------------------------------------------------------------------------------------------------------------------------
ESTADÍSTICAS GENERALES:
------------------------------------------------------------------------------------------------------------------------
  • Reducción total de regiones:        -16.8%
  • Coherencia promedio Classic RG:     95.8%
  • Coherencia promedio MGRG:           88.4%
  • Mejora en coherencia:               -7.4%
========================================================================================================================


INSIGHTS:
  MGRG tiene 88.4% de coherencia
  Classic RG tiene 95.8% de coherencia
  Mejora: -7.4%

# Visualización comparativa completa
fig = visualize_method_comparison(
    zones_data=zones_data,
    ndvi_results=ndvi_results,
    classic_results=segmentation_results,
    classic_stress=classic_stress_classification,
    mgrg_results=mgrg_results,
    mgrg_stress=mgrg_stress_classification,
    output_path=OUTPUT_DIR / 'complete_comparison_classic_vs_mgrg.png',
    dpi=300,
    show_plot=True
)

print("Visualización comparativa completa guardada")

Visualización comparativa guardada en: ..\..\img\results\classification\complete_comparison_classic_vs_mgrg.png

Visualización comparativa completa guardada

# Generar gráficos de comparativa A/B
fig, summary = visualize_ab_comparison_metrics(
    zones_data=zones_data,
    classic_results=segmentation_results,
    classic_quality=classic_quality_metrics,
    mgrg_results=mgrg_results,
    mgrg_quality=mgrg_quality_metrics,
    output_path=OUTPUT_DIR / 'ab_comparison_metrics.png',
    dpi=300,
    show_plot=True
)

# Acceder a las estadísticas de resumen
print(f"\nEstadísticas clave:")
print(f"  Reducción promedio: {summary['avg_reduction']:.1f}%")
print(f"  Mejora en coherencia: {summary['avg_coherence_improvement']:+.1f}%")
print(f"  MGRG es {summary['time_factor']:.1f}x más lento que Classic RG")



print("Gráficos de comparativa A/B guardados")

Gráficos de comparativa A/B guardados en: ..\..\img\results\classification\ab_comparison_metrics.png

================================================================================
RESUMEN DE MEJORAS PROMEDIO:
================================================================================
  • Reducción de regiones:      0.5%
  • Mejora en coherencia:       -7.4 puntos porcentuales
  • Tiempo Classic RG:          2.58s
  • Tiempo MGRG:                2.91s
  • Factor de tiempo:           1.1x
================================================================================


Estadísticas clave:
  Reducción promedio: 0.5%
  Mejora en coherencia: -7.4%
  MGRG es 1.1x más lento que Classic RG
Gráficos de comparativa A/B guardados

print("VALIDACIÓN CON DYNAMIC WORLD")

# Configuración de paths
DYNAMIC_WORLD_PATH = Path('../../data/dynamic_world')
RESULTS_PATH = OUTPUT_DIR / 'validation'
RESULTS_PATH.mkdir(parents=True, exist_ok=True)

VALIDACIÓN CON DYNAMIC WORLD

print_validation_classes()

Clases de validación:
  0: Water
  1: Crop
  2: Urban
  3: Bare Soil
  4: Other

# Mapeo de zonas
zone_mapping = {
    'mexicali': 'mexicali',
    'bajio': 'bajio',
    'sinaloa': 'sinaloa'
}

# Validar todas las zonas
validation_results = validate_all_zones(
    zones_data=zones_data,
    zone_mapping=zone_mapping,
    segmentation_results=segmentation_results,
    mgrg_results=mgrg_results,
    ndvi_results=ndvi_results,
    dynamic_world_path=DYNAMIC_WORLD_PATH,
    results_path=RESULTS_PATH,
    num_classes=len(VALIDATION_CLASSES),
    save_confusion_matrices=True
)

================================================================================
VALIDANDO: MEXICALI
================================================================================
  Classic RG: (1124, 922), 300 regiones
  MGRG: (1124, 922), 137 regiones
  Ground Truth: Dynamic World (real)
  Ground Truth redimensionado: (1124, 922)
  Ground Truth shape: (1124, 922)

  Mapeando segmentaciones a clases de validación...

======================================================================
METRICAS DE VALIDACION - MEXICALI
======================================================================

Mejora de MGRG sobre Classic RG: +1.1%

  ✓ Confusion matrices guardadas: ..\..\img\results\classification\validation\confusion_matrices_mexicali.png

================================================================================
VALIDANDO: BAJIO
================================================================================
  Classic RG: (1092, 1057), 451 regiones
  MGRG: (1092, 1057), 297 regiones
  Ground Truth: Dynamic World (real)
  Ground Truth redimensionado: (1092, 1057)
  Ground Truth shape: (1092, 1057)

  Mapeando segmentaciones a clases de validación...

======================================================================
METRICAS DE VALIDACION - BAJIO
======================================================================

Mejora de MGRG sobre Classic RG: +0.7%

  ✓ Confusion matrices guardadas: ..\..\img\results\classification\validation\confusion_matrices_bajio.png

================================================================================
VALIDANDO: SINALOA
================================================================================
  Classic RG: (1090, 1031), 632 regiones
  MGRG: (1090, 1031), 1181 regiones
  Ground Truth: Dynamic World (real)
  Ground Truth redimensionado: (1090, 1031)
  Ground Truth shape: (1090, 1031)

  Mapeando segmentaciones a clases de validación...

======================================================================
METRICAS DE VALIDACION - SINALOA
======================================================================

Mejora de MGRG sobre Classic RG: +33.7%

  ✓ Confusion matrices guardadas: ..\..\img\results\classification\validation\confusion_matrices_sinaloa.png

================================================================================
✓ VALIDACIÓN COMPLETADA PARA TODAS LAS ZONAS
================================================================================

# Calcular todas las métricas agregadas (TODO EN UNA FUNCIÓN)
df_aggregated, df_improvements, df_zone_stats = calculate_aggregate_metrics(
    validation_results=validation_results,
    use_display=True
)

====================================================================================================
MÉTRICAS AGREGADAS (Promedio ± Desviación Estándar)
====================================================================================================

====================================================================================================
ANÁLISIS DE MEJORA (MGRG vs Classic RG)
====================================================================================================

====================================================================================================
ESTADÍSTICAS DETALLADAS POR ZONA
====================================================================================================

====================================================================================================
RESUMEN FINAL DE VALIDACIÓN
====================================================================================================

Resultados en 3 zonas agrícolas:

Classic RG:
  • mIoU: 0.0168 ± 0.0124
  • Regiones promedio: 461

MGRG:
  • mIoU: 0.0205 ± 0.0173
  • Regiones promedio: 538

Mejora general: +22.3%
✓ MGRG supera a Classic RG en mIoU promedio
====================================================================================================

# Análisis completo de mejoras relativas (TODO EN UNA FUNCIÓN)
df_comparison, df_stats, improvements = analyze_relative_improvements(
    validation_results=validation_results,
    results_path=RESULTS_PATH,
    use_display=True
)

# Acceder a resultados
print(f"\nResumen:")
print(f"  MGRG es superior en {improvements['positive_count']} métricas")
print(f"  Mejora promedio: {improvements['avg_improvement']:+.1f}%")

# Exportar resultados
df_comparison.to_csv(RESULTS_PATH / 'relative_improvements.csv', index=False)
if df_stats is not None:
    df_stats.to_csv(RESULTS_PATH / 'statistical_tests.csv', index=False)

print("\nAnálisis de mejoras completado")

====================================================================================================
ANÁLISIS DE MEJORA RELATIVA: MGRG vs Classic RG
====================================================================================================

====================================================================================================
ANÁLISIS ESTADÍSTICO (Prueba t pareada)
====================================================================================================

Interpretación:
  *** p<0.01: Altamente significativo
  **  p<0.05: Significativo
  *   p<0.10: Marginalmente significativo
  ns: No significativo

====================================================================================================
VISUALIZACIÓN DE MEJORAS POR MÉTRICA
====================================================================================================

Gráfico guardado en: ..\..\img\results\classification\validation\relative_improvement_by_metric.png

====================================================================================================
RESUMEN DE MEJORAS
====================================================================================================

Métricas analizadas: 6
  • MGRG superior: 5 métricas (83%)
  • Classic RG superior: 1 métricas (17%)

Mejor mejora:
  • Recall: +38.3%

Peor resultado:
  • Precision: -8.2%

Mejora promedio general: +21.2%
====================================================================================================


Resumen:
  MGRG es superior en 5 métricas
  Mejora promedio: +21.2%

Análisis de mejoras completado

print('CLASIFICACIÓN SEMÁNTICA DE REGIONES')

# Define 6 land cover classes (bilingual)
print("\nLand cover classes:")
for class_id, class_name in LAND_COVER_CLASSES.items():
    print(f"  {class_id}: {class_name}")

CLASIFICACIÓN SEMÁNTICA DE REGIONES

Land cover classes:
  0: Water (Agua)
  1: Urban (Urbano)
  2: Bare Soil (Suelo Desnudo)
  3: Vigorous Crop (Cultivo Vigoroso)
  4: Stressed Crop (Cultivo Estresado)
  5: Grass/Shrub (Pasto/Arbustos)

fig = analyze_ndvi_zones(
    ndvi_results=ndvi_results,
    zones_data=zones_data,
    zone_mapping=zone_mapping,
    output_path=OUTPUT_DIR / 'ndvi_distribution_analysis.png',
    dpi=300
)

print("\nAnálisis de NDVI completado")

Valle de Mexicali:
  Min:  -0.580
  Max:  0.913
  Mean: 0.191
  Std:  0.141

El Bajío:
  Min:  -0.381
  Max:  0.888
  Mean: 0.315
  Std:  0.132

Valle de Culiacán:
  Min:  -0.493
  Max:  0.853
  Mean: 0.198
  Std:  0.195
Visualización NDVI guardada en: ..\..\img\results\classification\ndvi_distribution_analysis.png

Análisis de NDVI completado

# Clasificación completa de todas las zonas (TODO EN UNA FUNCIÓN)
semantic_results, df_all_classes, df_all_classes_formatted = classify_all_zones(
    mgrg_results=mgrg_results,
    ndvi_results=ndvi_results,
    embeddings_data=embeddings_data,
    zones_data=zones_data,
    zone_mapping=zone_mapping,
    resolution=10.0,
    min_size=50
)

# Mostrar DataFrame formateado
display(df_all_classes_formatted)

CLASIFICACIÓN SEMÁNTICA ZERO-SHOT

# Visualización completa de mapas semánticos (TODO EN UNA FUNCIÓN)
fig, colored_maps = visualize_semantic_maps(
    semantic_results=semantic_results,
    ndvi_results=ndvi_results,
    mgrg_results=mgrg_results,
    embeddings_data=embeddings_data,
    zones_data=zones_data,
    zone_mapping=zone_mapping,
    output_path=OUTPUT_DIR / 'semantic_maps_all_zones.png',
    dpi=300
)
print("Visualización de mapas semánticos completada")

VISUALIZACIÓN DE MAPAS SEMÁNTICOS

Valle de Mexicali:
  Shape: (1124, 922)
  Clases únicas: [0 1 2 3 4 5]
El Bajío:
  Shape: (1092, 1057)
  Clases únicas: [0 1 2 3 4]
Valle de Culiacán:
  Shape: (1090, 1031)
  Clases únicas: [0 1 2 3 4]

Visualización guardada en: ..\..\img\results\classification\semantic_maps_all_zones.png

Visualización de mapas semánticos completada

# Visualización de estadísticas por clase (TODO EN UNA FUNCIÓN)
figures = visualize_class_statistics(
    semantic_results=semantic_results,
    zones_data=zones_data,
    zone_mapping=zone_mapping,
    output_dir=OUTPUT_DIR,
    dpi=300
)
print("Visualización de estadísticas por clase completada")

VISUALIZACIÓN DE ESTADÍSTICAS POR CLASE (POR ZONA)

Estadísticas guardadas: ..\..\img\results\classification\class_statistics_plots_mexicali.png

Estadísticas guardadas: ..\..\img\results\classification\class_statistics_plots_bajio.png

Estadísticas guardadas: ..\..\img\results\classification\class_statistics_plots_sinaloa.png

Visualización completada para 3 zonas
Visualización de estadísticas por clase completada

print("RESUMEN AGREGADO POR CLASE")

# Agrupar por clase
class_summary = df_all_classes.groupby('Clase').agg({
    'Regiones': 'sum',
    'Área (ha)': 'sum',
    '% Área': 'mean',
    'NDVI Promedio': 'mean',
    'NDVI Desv. Est.': 'mean'
}).reset_index()

# Formatear
class_summary_formatted = class_summary.copy()
class_summary_formatted['Regiones'] = class_summary_formatted['Regiones'].astype(int)
class_summary_formatted['Área (ha)'] = class_summary_formatted['Área (ha)'].apply(lambda x: f"{x:.1f}")
class_summary_formatted['% Área'] = class_summary_formatted['% Área'].apply(lambda x: f"{x:.1f}%")
class_summary_formatted['NDVI Promedio'] = class_summary_formatted['NDVI Promedio'].apply(lambda x: f"{x:.3f}")
class_summary_formatted['NDVI Desv. Est.'] = class_summary_formatted['NDVI Desv. Est.'].apply(lambda x: f"±{x:.3f}")

display(class_summary_formatted)

RESUMEN AGREGADO POR CLASE

# Crear resumen agrupado por clase
class_summary = df_all_classes.groupby('Clase').agg({
    'Área (ha)': 'sum',
    'Regiones': 'sum'
}).reset_index()

# Generar visualización
print("\nGenerando gráfico de distribución de clases...")
fig = visualize_class_distribution_summary(
    class_summary=class_summary,
    output_path=OUTPUT_DIR / 'class_distribution_summary.png',
    dpi=300
)
print("\n✓ Gráfico de distribución de clases completado")

Generando gráfico de distribución de clases...

Generando gráfico de distribución de clases...

Gráfico de distribución guardado en: ..\..\img\results\classification\class_distribution_summary.png

✓ Gráfico de distribución de clases completado

# Visualización de estadísticas por clase (todas las zonas)
import sys
import importlib

# Reload module
if 'src.utils.visualization_utils' in sys.modules:
    importlib.reload(sys.modules['src.utils.visualization_utils'])

from src.utils.visualization_utils import visualize_class_statistics_all_zones
from src.classification.zero_shot_classifier import LAND_COVER_CLASSES, CLASS_COLORS
import numpy as np

# Prepare class names and colors
class_names = list(LAND_COVER_CLASSES.values())
colors_normalized = [np.array(CLASS_COLORS[i]) / 255.0 for i in range(6)]

# Generate visualization
fig = visualize_class_statistics_all_zones(
    semantic_results=semantic_results,
    zones_data=zones_data,
    zone_mapping=zone_mapping,
    class_names=class_names,
    colors_normalized=colors_normalized,
    output_path=OUTPUT_DIR / 'class_statistics_plots_all_zones.png',
    dpi=300
)

print("\n✓ Visualización de estadísticas por clase completada")

GENERANDO GRÁFICOS DE ESTADÍSTICAS

Gráficos guardados en: ..\..\img\results\classification\class_statistics_plots_all_zones.png

✓ Visualización de estadísticas por clase completada

# Run cross-validation
dw_results, all_agreements = cross_validate_with_dynamic_world(
    semantic_results=semantic_results,
    zones_data=zones_data,
    zone_mapping=zone_mapping,
    colored_maps=colored_maps,
    dynamic_world_path=Path('../../data/dynamic_world'),
    target_agreement=0.70
)

print("\n✓ Cross-validation con Dynamic World completada")

CROSS-VALIDATION CON DYNAMIC WORLD

Procesando: Valle de Mexicali
  Dynamic World shape: (2227, 2227)
  Unique classes: [0 1 2 3 4 5 6 7 8]
  Resized Dynamic World to: (1124, 922)

  === Resultados de Cross-Validation ===
  Water (Agua)                       : 3.4% agreement
  Urban (Urbano)                     : 0.0% agreement
  Bare Soil (Suelo Desnudo)          : 54.6% agreement
  Vigorous Crop (Cultivo Vigoroso)   : 1.7% agreement
  Stressed Crop (Cultivo Estresado)  : 16.2% agreement
  Grass/Shrub (Pasto/Arbustos)       : 0.2% agreement

  Overall Agreement: 50.7%
  Target (>70%): ⚠️ Below target

Procesando: El Bajío
  Dynamic World shape: (2228, 2227)
  Unique classes: [0 1 2 3 4 5 6 7 8]
  Resized Dynamic World to: (1092, 1057)

  === Resultados de Cross-Validation ===
  Water (Agua)                       : 1.9% agreement
  Urban (Urbano)                     : 32.3% agreement
  Bare Soil (Suelo Desnudo)          : 1.0% agreement
  Vigorous Crop (Cultivo Vigoroso)   : 85.9% agreement
  Stressed Crop (Cultivo Estresado)  : 63.9% agreement
  Grass/Shrub (Pasto/Arbustos)       : 0.0% agreement

  Overall Agreement: 51.8%
  Target (>70%): ⚠️ Below target

Procesando: Valle de Culiacán
  Dynamic World shape: (2227, 2228)
  Unique classes: [0 1 2 3 4 5 6 7 8]
  Resized Dynamic World to: (1090, 1031)

  === Resultados de Cross-Validation ===
  Water (Agua)                       : 2.3% agreement
  Urban (Urbano)                     : 22.1% agreement
  Bare Soil (Suelo Desnudo)          : 4.8% agreement
  Vigorous Crop (Cultivo Vigoroso)   : 0.0% agreement
  Stressed Crop (Cultivo Estresado)  : 26.7% agreement
  Grass/Shrub (Pasto/Arbustos)       : 0.0% agreement

  Overall Agreement: 11.2%
  Target (>70%): ⚠️ Below target


================================================================================
PROMEDIO DE ACUERDO EN TODAS LAS ZONAS: 37.9%
================================================================================


✓ Cross-validation con Dynamic World completada

# Visualización de comparación con Dynamic World
# Define Dynamic World color mapping
dw_colors = {
    0: (65, 155, 223),   # Water (Agua) - Blue
    1: (57, 125, 73),    # Trees (Árboles) - Dark Green
    2: (136, 176, 83),   # Grass (Pasto) - Light Green
    3: (122, 135, 198),  # Flooded Vegetation (Vegetación Inundada) - Purple
    4: (228, 150, 53),   # Crops (Cultivos) - Orange
    5: (223, 195, 90),   # Shrub (Arbustos) - Yellow
    6: (196, 40, 27),    # Built (Construido) - Red
    7: (165, 155, 143),  # Bare (Suelo Desnudo) - Gray
    8: (179, 159, 225),  # Snow/Ice (Nieve/Hielo) - Light Purple
}

# Map our classes to Dynamic World classes
our_to_dw = {
    0: 0,  # Water -> Water
    1: 6,  # Urban -> Built
    2: 7,  # Bare Soil -> Bare
    3: 4,  # Vigorous Crop -> Crops
    4: 4,  # Stressed Crop -> Crops
    5: 2,  # Grass/Shrub -> Grass
}

# Generate comparison visualizations
comparison_figures = visualize_dynamic_world_comparison(
    dw_results=dw_results,
    semantic_results=semantic_results,
    zones_data=zones_data,
    zone_mapping=zone_mapping,
    our_to_dw_mapping=our_to_dw,
    dw_colors=dw_colors,
    output_dir=OUTPUT_DIR,
    dpi=300
)

print("\n✓ Comparación con Dynamic World completada")

GENERANDO VISUALIZACIONES DE COMPARACIÓN

  ✓ Comparación guardada: dynamic_world_comparison_mexicali.png

  ✓ Comparación guardada: dynamic_world_comparison_bajio.png

  ✓ Comparación guardada: dynamic_world_comparison_sinaloa.png

✓ Visualizaciones completadas para 3 zonas

✓ Comparación con Dynamic World completada

# ANÁLISIS DE ESTRÉS EN CULTIVOS (TODAS LAS ZONAS)
import sys
import importlib

# Reload module
if 'src.utils.classification_utils' in sys.modules:
    importlib.reload(sys.modules['src.utils.classification_utils'])

from src.utils.classification_utils import analyze_crop_stress

# Analyze crop stress across all zones
df_stress_raw, df_stress_formatted, df_total_raw, df_total_formatted = analyze_crop_stress(
    semantic_results=semantic_results,
    zones_data=zones_data,
    zone_mapping=zone_mapping
)

# Display results per zone
print("\nAnálisis de Estrés en Cultivos por Zona:")
display(df_stress_formatted)

# Display aggregated totals
print("\nTotales Agregados (Todas las Zonas):")
display(df_total_formatted)

Análisis de Estrés en Cultivos por Zona:

Totales Agregados (Todas las Zonas):

# Display optimal thresholds
print("\n" + "="*70)
print("OPTIMAL THRESHOLDS (from experimental analysis)")
print("="*70)
print("\nClassic RG:")
print("  Threshold: 0.1 (NDVI difference)")
print("  Rationale: Balance between detail and coherence")
print("\nMGRG:")
print("  Threshold: 0.95 (cosine similarity)")
print("  Rationale: Optimal class separation without fragmentation")
print("="*70)

======================================================================
OPTIMAL THRESHOLDS (from experimental analysis)
======================================================================

Classic RG:
  Threshold: 0.1 (NDVI difference)
  Rationale: Balance between detail and coherence

MGRG:
  Threshold: 0.95 (cosine similarity)
  Rationale: Optimal class separation without fragmentation
======================================================================

python scripts/analyze_region.py \
  --bbox "32.45,-115.35,32.55,-115.25" \
  --date "2024-01-15" \
  --output "output/mexicali" \
  --threshold 0.95 \
  --formats json,png

Nombre	Matrícula	Rol	Contribuciones Principales	Horas
Carlos Aaron Bocanegra Buitron	A01796345	FullStack Developer	FastAPI backend, RG Clásico, Pipeline end-to-end, Integración de componentes	30h
Arthur Jafed Zizumbo Velasco	A01796363	ML Engineer, Tech Lead & Prithvi Integration	Prithvi integration, MGRG implementation, Validación con Dynamic World, Clasificación semántica, Comparativas A/B	70h
Luis Santiago Vázquez Mancilla	A01796029	Front Developer & Visualization	Nuxt 3 frontend, Visualizaciones profesionales, MapLibre integration	20h
Edgar Oviedo Navarro	A01795260	Product Owner & Documentation Lead	Artículo científico, Video presentación, Documentación técnica, Teoría y fundamentos	20h

Zona	Método	Regiones	Coherencia	mIoU	Mejora
Mexicali	Classic RG	207	72.5%	0.1123	-
	MGRG	24	94.2%	0.1224	+9%
Bajío	Classic RG	775	65.3%	0.0020	-
	MGRG	10	98.1%	0.1544	+7620%
Sinaloa	Classic RG	934	68.7%	0.0004	-
	MGRG	19	96.8%	0.1278	+31850%
Promedio	Classic RG	639	68.8%	0.0382	-
	MGRG	18	96.4%	0.1349	+252.8%

	name	bbox	description	crops	state	climate	irrigation
Zona ID
mexicali	Valle de Mexicali	[32.4, -115.4, 32.5, -115.3]	Valle de Mexicali - Agricultura intensiva de riego	Trigo, algodón, alfalfa	Baja California	Árido	Intensivo
bajio	El Bajío	[20.8, -101.5, 20.9, -101.4]	El Bajío - Región agrícola diversificada	Sorgo, maíz, hortalizas	Guanajuato	Templado	Mixto
sinaloa	Valle de Culiacán	[24.7, -107.5, 24.8, -107.4]	Valle de Culiacán - Agricultura de exportación	Tomate, chile, maíz	Sinaloa	Tropical	Tecnificado

	Zona	Regiones	Coherencia (%)	Cobertura (%)	Tamaño Promedio (px)	Tamaño Promedio (ha)	Rango Tamaño (px)	Desv. Est. Tamaño
0	Valle de Mexicali	300	95.7	85.0	2935	29.35	50 - 718821	41408
1	El Bajío	451	96.5	79.9	2046	20.46	50 - 447724	22006
2	Valle de Culiacán	632	95.2	55.8	993	9.93	50 - 219401	9714

	Zona	Regiones	Coherencia (%)	Cobertura (%)	Tamaño Promedio (px)	Tamaño Promedio (ha)	Rango Tamaño (px)	Desv. Est. Tamaño
0	Valle de Mexicali	137	88.9	98.6	7459	74.59	51 - 747357	63579
1	El Bajío	297	88.9	95.8	3722	37.22	58 - 887308	51417
2	Valle de Culiacán	1181	87.3	89.0	846	8.46	51 - 99084	3446

	Zona	Total Regiones	Alto	Alto (%)	Medio	Medio (%)	Bajo	Bajo (%)
0	Valle de Mexicali	300	142	47.3	107	35.7	51	17.0
1	El Bajío	451	241	53.4	123	27.3	87	19.3
2	Valle de Culiacán	632	356	56.3	167	26.4	109	17.2

	Zona	Classic RG Regiones	Classic RG Cobertura (%)	Classic RG Tiempo (s)	MGRG Regiones	MGRG Cobertura (%)	MGRG Tiempo (s)	Reducción (%)
0	Valle de Mexicali	300	85.0	2.62	137	98.6	2.87	54.3
1	El Bajío	451	79.9	2.91	297	95.8	3.38	34.1
2	Valle de Culiacán	632	55.8	2.20	1181	89.0	2.47	-86.9

	Zona	Total Regiones	Alto	Alto (%)	Medio	Medio (%)	Bajo	Bajo (%)
0	Valle de Mexicali	137	95	69.3	27	19.7	15	10.9
1	El Bajío	297	263	88.6	29	9.8	5	1.7
2	Valle de Culiacán	1181	1051	89.0	121	10.2	9	0.8

	Metodo	Regiones	mIoU	Weighted mIoU	F1-Score	Precision	Recall	Pixel Acc
0	Classic RG	300	0.0029	0.0001	0.0029	0.0016	0.2218	0.0051
1	MGRG	137	0.0030	0.0001	0.0030	0.0016	0.2446	0.0016

	Metodo	Regiones	mIoU	Weighted mIoU	F1-Score	Precision	Recall	Pixel Acc
0	Classic RG	451	0.0144	0.0011	0.0140	0.0073	0.1653	0.0192
1	MGRG	297	0.0145	0.0011	0.0141	0.0073	0.2269	0.0242

	Metodo	Regiones	mIoU	Weighted mIoU	F1-Score	Precision	Recall	Pixel Acc
0	Classic RG	632	0.0330	0.0160	0.0309	0.0367	0.0955	0.0237
1	MGRG	1181	0.0441	0.0214	0.0405	0.0330	0.1962	0.0332

	Método	mIoU	Weighted mIoU	Macro F1	Precision	Recall	Pixel Acc	Regiones
0	Classic RG	0.0168 ± 0.0124	0.0057 ± 0.0073	0.0159 ± 0.0115	0.0152 ± 0.0154	0.1609 ± 0.0517	0.0160 ± 0.0079	461 ± 136
1	MGRG	0.0205 ± 0.0173	0.0075 ± 0.0098	0.0192 ± 0.0157	0.0140 ± 0.0137	0.2226 ± 0.0200	0.0197 ± 0.0133	538 ± 459

	Métrica	Mejora Media	Desv. Est.	Rango
0	mIoU	+11.8%	±15.4%	[+0.7%, +33.7%]
1	Weighted mIoU	+11.9%	±15.4%	[+0.1%, +33.6%]
2	Macro F1	+10.9%	±14.2%	[+0.8%, +31.0%]
3	Precision	-3.2%	±4.9%	[-10.1%, +1.1%]
4	Recall	+51.0%	±40.0%	[+10.3%, +105.4%]
5	Pixel Acc	-0.5%	±48.1%	[-68.0%, +40.0%]

	Zona	Classic Regiones	MGRG Regiones	Reducción	Classic mIoU	MGRG mIoU	Mejora mIoU	Ground Truth
0	Mexicali	300	137	54.3%	0.0029	0.0030	+1.1%	Real
1	Bajio	451	297	34.1%	0.0144	0.0145	+0.7%	Real
2	Sinaloa	632	1181	-86.9%	0.0330	0.0441	+33.7%	Real

	Métrica	t-statistic	p-value	Significancia
0	mIoU	1.019	0.4155	ns (no significativo)
1	Weighted mIoU	1.001	0.4223	ns (no significativo)
2	Macro F1	1.022	0.4144	ns (no significativo)
3	Precision	-1.009	0.4194	ns (no significativo)
4	Recall	2.746	0.1109	ns (no significativo)
5	Pixel Accuracy	0.973	0.4332	ns (no significativo)

	Clase	Regiones	Área (ha)	% Área	NDVI Promedio	NDVI Desv. Est.
0	Bare Soil (Suelo Desnudo)	1131	16928.6	55.2%	0.181	±0.051
1	Grass/Shrub (Pasto/Arbustos)	6	87.9	0.9%	0.629	±0.026
2	Stressed Crop (Cultivo Estresado)	195	12222.9	37.8%	0.391	±0.072
3	Urban (Urbano)	243	1646.3	5.4%	0.062	±0.029
4	Vigorous Crop (Cultivo Vigoroso)	5	11.0	0.0%	0.698	±0.012
5	Water (Agua)	35	374.7	1.2%	0.054	±0.016

	Nivel de Estrés	Regiones	Área Total (ha)	% del Total
0	Low Stress	18	206.0	1.7%
1	Medium Stress	47	1814.1	14.8%
2	High Stress	130	10202.8	83.5%

	Zona	Clase	Regiones	Área (ha)	% Área	NDVI Promedio	NDVI Desv. Est.
0	Valle de Mexicali	Water (Agua)	4	18.1	0.2%	0.087	±0.015
1	Valle de Mexicali	Urban (Urbano)	5	20.2	0.2%	0.065	±0.026
2	Valle de Mexicali	Bare Soil (Suelo Desnudo)	86	9412.5	92.1%	0.165	±0.051
3	Valle de Mexicali	Vigorous Crop (Cultivo Vigoroso)	1	1.8	0.0%	0.734	±0.000
4	Valle de Mexicali	Stressed Crop (Cultivo Estresado)	35	678.9	6.6%	0.430	±0.077
5	Valle de Mexicali	Grass/Shrub (Pasto/Arbustos)	6	87.9	0.9%	0.629	±0.026
6	El Bajío	Water (Agua)	1	8.8	0.1%	0.058	±0.000
7	El Bajío	Urban (Urbano)	38	170.6	1.5%	0.060	±0.030
8	El Bajío	Bare Soil (Suelo Desnudo)	224	1637.5	14.8%	0.204	±0.053
9	El Bajío	Vigorous Crop (Cultivo Vigoroso)	3	8.2	0.1%	0.739	±0.036
10	El Bajío	Stressed Crop (Cultivo Estresado)	31	9230.2	83.5%	0.367	±0.079
11	Valle de Culiacán	Water (Agua)	30	347.9	3.5%	0.016	±0.032
12	Valle de Culiacán	Urban (Urbano)	200	1455.5	14.6%	0.060	±0.031
13	Valle de Culiacán	Bare Soil (Suelo Desnudo)	821	5878.5	58.8%	0.172	±0.049
14	Valle de Culiacán	Vigorous Crop (Cultivo Vigoroso)	1	1.0	0.0%	0.620	±0.000
15	Valle de Culiacán	Stressed Crop (Cultivo Estresado)	129	2313.8	23.1%	0.376	±0.062

	Zona	Nivel de Estrés	Regiones	Área (ha)	% del Total de Cultivos
0	Valle de Mexicali	Low Stress	8	146.5	21.6%
1	Valle de Mexicali	Medium Stress	13	328.4	48.4%
2	Valle de Mexicali	High Stress	14	203.9	30.0%
3	El Bajío	Low Stress	2	4.0	0.0%
4	El Bajío	Medium Stress	5	65.1	0.7%
5	El Bajío	High Stress	24	9161.0	99.3%
6	Valle de Culiacán	Low Stress	8	55.4	2.4%
7	Valle de Culiacán	Medium Stress	29	1420.5	61.4%
8	Valle de Culiacán	High Stress	92	837.8	36.2%

Instituto Tecnológico y de Estudios Superiores de Monterrey¶

Maestría en Inteligencia Artificial Aplicada¶

Materia: Visión Computacional¶

Profesor: Gilberto Ochoa¶

Título: Segmentación Semántica Basada en Region Growing Aplicada a la Percepción Remota Agrícola¶

Equipo 24 - Contribuciones¶

Tabla de Contenidos¶

1. Resumen Ejecutivo¶

1.1 Resumen (Español)¶

1.2 Abstract (English)¶

2. Introducción¶

3. 🔬 Estado del Arte¶

3.1 Region Growing Clásico¶

3.2 Foundation Models en Percepción Remota¶

Prithvi-EO-1.0-100M (NASA/IBM, 2024)¶

SatMAE (Cong et al., 2022)¶

3.3 Hibridación DL-OBIA¶

CRGNet (Ghamisi et al., 2022)¶

3.4 Nuestra Propuesta: MGRG¶

3.5 Referencias Principales¶

4. 📐 Fundamentos Teóricos¶

4.1 Algoritmo Region Growing¶

4.2 Índices Espectrales¶

NDVI (Normalized Difference Vegetation Index)¶

NDWI (Normalized Difference Water Index)¶

4.3 Embeddings Semánticos¶

4.4 Cosine Similarity¶

4.5 Métricas de Evaluación¶

IoU (Intersection over Union)¶

Mean IoU¶

F1-Score¶

Coherencia Espacial¶

5. Setup e Instalación¶

5.1. Evidencia: pyproject.toml del Proyecto¶

5.2 Verificación de GPU¶

5.3 Imports Organizados¶

6. Descarga y Preparación de Datos¶

6.1 Configuración de Zonas Mexicanas¶

6.2 Configuración de Fecha y Parámetros¶

6.3 Descarga y Procesamiento de Imágenes HLS¶

6.4 Visualización de Imágenes RGB Originales¶

7. Método 1: Region Growing Clásico¶

7.1 Implementación del Algoritmo¶

7.2 Cálculo NDVI¶

7.3 Generación de Semillas¶

7.4 Segmentación¶

7.5 Clasificación por Nivel de Estrés¶

7.6 Métricas de calidad¶

7.7 Resumen Comparativo¶

7.8 Visualización de Resultados¶

8. Método 2: MGRG (Region Growing Semántico)¶

8.1 Extracción de Embeddings con Prithvi¶

8.2 Carga de Embeddings¶

8.2 Implementación MGRG¶

8.3 Clasificación por Estrés¶

8.4 Métricas de Calidad¶

8.5 Resumen Comparativo¶

8.6 Visualización de Resultados¶

9. Comparativa entre Classic RG vs MGRG¶

9.1 Resumen Comparativo¶

9.2 Comparativa de Visualizaciones¶

10. Métricas Cuantitativas¶

Métricas Implementadas¶

10.1 Validación Cuantitativa con Reference Dataset (Dynamic World)¶

10.2 Mapeo de clases¶

10.3 Validación en todas las zonas¶

10.4 Tabla comparativa resultados¶

10.5 Análisis de Mejora Relativa¶

11. Clasificación Semántica¶

11.1 Clases¶

11.2 Exploración inicial¶

11.3 Clasificación Zero-Shot¶

11.4 Visualización de Mapas Semánticos¶

11.6 Gráficos¶

11.7 Resumen agregado por clase (promedio de todas las zonas)¶

11.8 Gráfico de distribución de clases¶

11.9 Visualizaciones de Estadísticas por clase¶

12. Cross-Validation con Dynamic World (DW)¶

12.1 Resultados de Cross-Validation con DW¶

12.2 Visualización Comparativa con Dynamic World¶